QoS guarantee in self-backhauled LTE mesh networks

(1)

QoS Guarantee in Self-Backhauled LTE Mesh Networks

Romain Favraud, Navid Nikaein and Chia-Yu Chang Communication Systems Department, EURECOM, France

Abstract—LTE is deployed in most countries and continuously evolving to match new use cases as well as requirements. While it is expected to evolve toward 5G deployments for outdoor and long range communications, it will also be used for Public Safety services in major countries. Among that, new scenarios call for wider networks relying on dynamic meshing of base stations.

Leveraging the LTE air interface for base station meshing is an appealing idea but is not straightforward to guarantee quality of service (QoS). In this article, we study and evaluate scheduling strategies for multi-hop LTE mesh networks relying on the LTE relay channel. We firstly present the LTE relay channel and the scheduling problem for in-band LTE self-backhauling. We then propose a practical cross-layer method in order to fulfill the QoS requirements of real-time flows in such network. We finally evaluate the proposed method through extensive simulations. We show the effectiveness of the proposed approach compared to a legacy method in terms of both QoS requirement satisfaction of real-time flows and throughput enhancement of elastic flows.

I. INTRODUCTION

Long Term Evolution (LTE) is now the 4G cellular network of reference. 3GPP LTE specifications are expanding its use- cases and increasing its features at every new release to become the 5G outdoor and long range radio access technology (RAT). Moreover, LTE is expected to be the next RAT for Public Safety (PS) communications and specifications items have emerged in this regard [1]. PS networks have several specific requirements including reliability and resiliency while demanding specific Quality of Service (QoS) requirements.

In common PS scenarios, LTE Base Stations (BSs), called eNodeBs (eNBs), may lose access to the Evolved Packet Core (EPC) due to some outage or specific deployments.

When that happens, they are unable to provide any service to their served users. This is an issue for PS networks and is addressed by 3GPP through the Isolated E-UTRAN concept that allows eNBs to continue providing minimal services for local PS users (TS 22.346, TR 23.797). In [2], we proposed an evolution of the Isolated E-UTRAN concept in the form of a new BS architecture for nomadic LTE networks called as enhanced eNB (e2NB). It embeds essential core network functions to ensure local services and has the ability to connect to other similar BSs leveraging the LTE relay interface to create a mesh network. Fig. 1 shows an example network topology. The e2NB can support new use cases where dynamic meshing among fix and/or moving BSs is highly required. In [3], we showed that the relay channel (calledUn) performance

Research leading to these results received funding from Naval Group and 5G-PPP COHERENT project under grant agreement 671639.

e2NB

e2NB e2NB

UE

Uu UE UE

eNB

μEPC

Coordination &

Orchestration Entity (COE) MMEHSS

TX/RX Radio Chain

vUE Un Un Uu

Fig. 1: LTE mesh network based on e2NB with LTE backhaul is close to the legacy LTE eNB ↔ User Equipment (UE) interface (calledUu) and that it allows efficient transmissions between e2NBs. However, one of the main challenges when meshing e2NBs, especially when reusing a single frequency band, is that the transmission, reception and resource allocation between BSs shall be coordinated to avoid excessive blocking due to interference and to guarantee the performance of end-to-end connections.

In this paper, we exploit the relay channels and propose the genuine resource allocation algorithm to evolve LTE from a single-hop network to a multi-hop meshing of several fixed and/or mobile BSs. We briefly recall the properties of the LTE relay physical channels and the inherent problem to reuse a single frequency band to mesh a wireless network in Section II.

In Section III, we present our target use cases and then state the associated problems, objectives and variables. In Section IV, we propose a cross-layer semi-centralized method to allocate resources of the network while taking into account the QoS requirements. In Section V, we evaluate the performance of the proposed method over several topologies via extensive simulations and we compare the results to other resource allocation algorithms. Finally, We conclude this work and present the next steps of our research.

II. BACKGROUND INFORMATION

Our considered problem is based on the LTE relay channel for in-band self-backhauling (i.e. single frequency networking) to mesh LTE BSs. Hence, we recall the background information and the inherent problem in this section.

A. LTE relay channel (Un)

LTE Relay is specified in 3GPP Release 10 allowing a relay node (RN) to serve UEs on its access link and reach the EPC through its backhaul link with its anchor eNB, called Donor

(2)

eNB (DeNB). The LTE relay interface is denoted as Un [4]

and is used to realize the in-band backhaul link between a LTE RN and its DeNB. As this link is in-band, it shares the same frequency band as the access link relying on a time division multiplexing mechanism. To ensure that the channel between RN and its UEs complies with the legacy access link interface (Uu), the Un interface relies on a property that is originally introduced in LTE eMBMS (enhanced Multimedia Broadcast Multicast Service, a.k.a. MBSFN) to differentiate multicast/broadcast subframes from the unicast ones. We recall that an LTE frame is made up of ten 1ms subframes (SFs) that carries a number of resource blocks (RBs) depends on the channel bandwidth. In an LTE FDD (Frequency Division Duplex) frame, a maximum of 6 MBSFN SFs are allowed. A relay can exploit properties of these MBSFN SFs to receive specifically formatted unicast RBs from its DeNB on the downlink (DL) channel instead of always being transmitting to its UEs on the access link [4]. This means that the maximum data rate on the Un interface is reduced to 60% of what can be achieved on theUu interface.

B. Single frequency wireless mesh network

While there are several works regarding the use of LTE Un interface for self-backhauling [5], [6]; to the best of our knowledge, none of them considered meshing BSs that host their own EPC and that do not have a unique backhaul path towards a gateway. As indicated hereinbefore, we advocate that the Un interface can be reused to extend the eNB capability with in-band self-backhauling to realize self-organized LTE mesh network among BSs [1], [2]. The e2NB incorporates virtual UEs (vUEs) that are used to initiate and maintain connections to the adjacent e2NBs using theUn interface.

However, realizing an efficient wireless mesh network on a single frequency is still an open research problem. As all nodes share the resources within the same frequency band, each transmitter will potentially become an interferer to others. Thus, collisions among transmissions may happen if the channel access is not coordinated properly, which in turn limits the achievable rate. Further, there exist more issues in wireless mesh networks as surveyed in [7], including (a) topology control, (b) routing, (c) link scheduling, (d) interference measurement and (e) power control. These aforementioned problems are highly inter-dependent across layers and transmitting nodes.

To deal with these issues jointly, we propose a coordinated and cross-layer approach to unleash the performance barrier when meshing e2NBs in a single frequency band. In that sense, the per-flow QoS can be guaranteed by a jointly consideration on the topology control, path routing, link scheduling, and power control. Our proposed approach is based on an architecture relying on a logically centralized control entity that manages and orchestrates the induced mesh network across medium access control (MAC) and network layer through policy enforcement, as described in [1], [2]. Based on the results of such centralized entity, the distributed entities will do local scheduling for both access and backhaul links.

Fig. 2: Example of MBSFN SFs use for inband backhauling III. PROBLEM STATEMENT AND OBJECTIVES

As mentioned in the previous section, it is necessary to define and enforce policies spanning across the relevant network layers and nodes to realize an efficient single frequency mesh network. We firstly define the objectives considering the use cases and then present the problem as well as its variables.

A. Traffic flows

In this work, we focus on public safety scenario with two types of traffic flow: real-time traffic with latency requirement (e.g. VoIP) and elastic traffic that is served in the best effort manner. As our considered scenario is extended from the Isolated E-UTRAN and each BS can provide services to other BSs through the self-backhaul LTE mesh network, hence the source and destination of traffic pattern can heterogeneously be composed of intra-cell and inter-cell UE-to-UE (U2U), eNB- to-eNB (e2e) and UE-to-eNB (U2e) communications. For the real-time traffic, we consider VoIP traffic with target maximum one-way-delay of 150ms for 95% of the packet to ensure a quality call with a Mean Opinion Score (MOS) of 3.5 using a G.729 codec [8]. For the elastic traffic, we consider multimedia data transfer with the objective to maximize its throughput.

B. Problem statement

We assume a single frequency FDD LTE network relying on e2NBs [3]. Each e2NB is equipped with a single omni directional antenna. Based on two entities (eNB, vUE(s)) and two interfaces (Un, Uu) in e2NB (see Fig. 1), the DL and UL SF allocation to either entity with corresponding interface is required as the example illustrated in Fig. 2 using all SFs (6 MBSFN SFs and 4 non-MBSFN SFs).

In our problem, we will allocate the (a) MBSFN SFs for backhaul transportation, (b) destination e2NB of backhaul relaying, and (c) transmission power. The resource allocation algorithm enforces policy to each e2NB on allocated transmission (Tx) and reception (Rx) SFs. Such policy depends on the traffic patterns on the mesh (which is a combination of U2U, U2e and e2e communications), on the diverse objectives (latency requirements and throughput maximization) of higher-layer applications, on the network topology, and on the selection of routing paths. Therefore, the main variables at the network level include:

∙ Tx/Rx SF allocation at each e2NB

∙ Power allocation of each SF at each eNB

∙ Traffic forwarding paths

During the allocation, a trade-off can be observed between allowing more e2NBs to communicate while in the mean time reducing the perceived interference by each node. Considering

(3)

the aforementioned two types of traffic flow, our approach has two objectives:

∙ Guarantee the latency of real-time flows (i.e. VoIP)

∙ Maximize the throughput of elastic flows IV. COORDINATED SCHEDULING

The proposed approach is based on a logically centralized coordinated and orchestration entity (COE) that schedules nodes and selects data forwarding paths by enforcing the policy at each BS (refer to Fig. 1). The COE follows a hierarchical design and is composed of a centralized entity that is connected to a number of COE agents [9], one per e2NB in a typical deployment. The COE agent can either act as a local COE with a limited network view handling control delegated by the centralized entity, or in coordination with other agents and the centralized entity. The communication between the centralized entity and agents is done through message ex- changes allowing a bi-directional interaction between them.

In one direction, the COE agent sends measured performance indicators and e2NB status to the centralized COE and other agents, while in the other direction the centralized entity enforces policies that define the operation to be executed by the agents and the underlying eNB and vUEs.

The aforementioned design provides the required flexibility in realizing the COE, and is able to reduce the control overhead by delegating more functions to the COE agent at the cost of less coordinated operation.

A. Topology Control

As we are using omni directional antennas, activating a link from a BS means interfering with all other links, regardless of the activated link. In order to remedy the interference from other links, we apply the power control and adaptive modulation and coding (AMC) scheme inUninterface that is already applied inUuinterface. To enable these two schemes, the received signal power and link quality shall be measured in the Uninterface.

Moreover, LTE allows for multi point-to-point transmissions from a single BS using OFDMA to transport unicast data to several nodes in the same slot instead of communication to each other nodes in the per-slot basis. Via utilizing such property, the scheduler at each BS is responsible to (a) select which neighboring BS(s) to transmit to and (b) pick the applied modulation and coding scheme (MCS) rather than applying topology-dependent policy that requires a priori topology information.

B. Global routing

We apply the Dijkstra’s algorithm at COE to find the shortest path to route traffic in backhaul. Such algorithm can minimize per-flow latency provided that the traffic load is below the maximum throughput to avoid congestion. Never- theless, a better performance can be achieved if an adaptive distributed mesh routing is used to cope with different traffic pattern and network topology [10].

Algorithm 1: Coordinated Scheduling Algorithm Input : 𝑃𝑆𝐹 is the SuperFrame update periodicity

𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡is the list of e2NBs begin

while𝑛𝑒𝑡𝑤𝑜𝑟𝑘𝑟𝑢𝑛𝑛𝑖𝑛𝑔do

𝑆𝐹 =𝑆𝐹+ 1/* current subframe identifier */

𝑇 𝑟𝑦𝑁𝑒𝑤= 0 ifeventthen

𝑇 𝑟𝑦𝑁𝑒𝑤= 1

if𝑆𝐹 ≡0 (mod𝑃𝑆𝐹)𝑜𝑟 𝑇 𝑟𝑦𝑁𝑒𝑤== 1then 𝑔𝑒𝑡 𝑚𝑒𝑡𝑟𝑖𝑐𝑠()

𝑐𝑜𝑚𝑝𝑢𝑡𝑒(𝐿𝑆𝐹)(c.f. Eq. 1) 𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑= 0

while𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑== 0do 𝑐𝑜𝑚𝑝𝑢𝑡𝑒(𝑆𝐹𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒)(c.f. Eq. 2) 𝑐𝑜𝑚𝑝𝑢𝑡𝑒(𝑆𝐹𝑒𝑙𝑎𝑠𝑡𝑖𝑐)(c.f. Eq. 3) [𝑆𝐹 𝑇 𝑋, 𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑] =

𝑐𝑒𝑛𝑡𝑟𝑎𝑙𝑖𝑧𝑒𝑑 𝑁𝑆(𝑆𝐹𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒, 𝑆𝐹𝑒𝑙𝑎𝑠𝑡𝑖𝑐) (c.f. Alg. 2)

𝐿𝑆𝐹 =𝐿𝑆𝐹+ 10

if𝐿𝑆𝐹 >𝑀𝑎𝑥𝐿𝑎𝑡∧𝑇 𝑟𝑦𝑁𝑒𝑤== 1then 𝑟𝑒𝑗𝑒𝑐𝑡(𝑓𝑙𝑜𝑤)/*Reject realtime flows*/

𝑐𝑜𝑚𝑝𝑢𝑡𝑒(𝐿𝑆𝐹)(c.f. Eq. 1) 𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑= 0 foreach𝑢∈𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡do

𝑙𝑜𝑐𝑎𝑙 𝐿𝑆(𝑆𝐹, 𝑆𝐹 𝑇 𝑋[𝑢])(c.f. Alg. 3)

C. Subframe scheduling

The main scheduling problem is to share the frequency (RBs) and time resources (MBSFN SFs) between e2NBs. Such scheduling is achieved at two different levels of different time- scales. Algorithm 1 provides a high-layer scheduling approach that summarizes the two different levels. A centralized node scheduler (NS) (Algorithm 2) is executed periodically or on reaction to an event (for instance, a new real-time flow with latency requirements) and it defines which MBSFN SFs shall be utilized by each e2NB for transmission. Then, each e2NB utilizes the local link scheduler (LS) that is executed at every SF to perform the per-link scheduling (Algorithm 3).

1) Centralized node scheduling: Without loss of generality, we only consider a time-domain resource allocation (i.e. SF allocation) in DL direction (i.e. from eNB to vUE) among the neighboring nodes for self-backhauling, as the UL SFs are allocated correspondingly to these allocated DL SFs. The proposed node scheduling determines the active SFs for each BS within a superframe to schedule the transmission and reception in accompany with the corresponding vUE(s). Its goal is to allocate enough SFs to each e2NB to fulfill the real- time traffic transportation. However, if that is not possible, Algorithm 1 guarantees that each e2NB gets at least one SF per superframe through iteration in the superframe length.

The resource allocation within a superframe for each e2NB is repeated until the node scheduling gets updated, which is triggered by traffic load and/or channel variabilities in the superframe. The length of the superframe is determined in such a way to satisfy the latency requirement for real-time traffic flows, and remains the same for all e2NBs in the network (generally in the order of tens of millisecond). It is

(4)

computed as a function of expected maximum number of hops, denoted as𝑀_{ℎ𝑜𝑝𝑠}in Eq. (1), that a given flow may experience.

𝐿_𝑆𝐹 =⌈𝑀𝑎𝑥𝐿𝑎𝑡/((𝑀_{ℎ𝑜𝑝𝑠}+𝑜𝑓𝑓𝑠𝑒𝑡))⌉ (1) The 𝑀𝑎𝑥𝐿𝑎𝑡represents the maximum acceptable latency for the real-time flows (e.g. 150ms for VoIP). The 𝑜𝑓𝑓𝑠𝑒𝑡is the stretch factor of𝑀ℎ𝑜𝑝𝑠, and it depends mainly on the mobility pattern (i.e. speed) and network size.

The algorithm relies on some a priori knowledge of real- time traffic flows in the network to proportionally allocate resources among BSs while respecting the QoS requirements.

Some real-time flow control policy can further be applied to avoid accepting too many flows that could overload the network. The following inputs are required for each link:

∙ Expected traffic load of real-time flows;

∙ Expected link quality or the resulted MCS for each flow;

∙ Utilization ratio of the allocated resources during the last node scheduling period.

The first metric, denoted as 𝑆𝐹_{𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒}[𝑢], is computed to determine the number of SFs during a superframe required by e2NB𝑢to transmit the real-time flows. It is the main control variable and is also used to sort e2NBs in the descending order before allocating the SFs. It is computed centrally as follows:

𝑆𝐹𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢] =

∑

𝑣∈𝑁(𝑢)𝑃 𝑅𝐵𝑢,𝑣(𝑡𝑝𝑡𝑢,𝑣×𝐿𝑆𝐹)

𝑁𝑃 𝑅𝐵 (2)

where 𝑡𝑝𝑡_𝑢,𝑣 is the sum of the throughput for real-time flows passing through link𝑢→𝑣(i.e. the transportation from node𝑢 to node𝑣).𝑃 𝑅𝐵_𝑢,𝑣(𝑥)returns the number of PRBs required to transmit 𝑥 bits given the current channel quality on link 𝑢→𝑣,𝑁(𝑢)is a set that comprises all neighboring nodes of 𝑢and𝑁𝑃 𝑅𝐵 is the total number of PRBs per SF.

A second metric is computed to increase the number of allocated SFs at e2NB 𝑢, denoted as 𝑆𝐹_{𝑒𝑙𝑎𝑠𝑡𝑖𝑐}[𝑢], for links that are saturated due to too many elastic flows passing through leading to a higher utilization ratio. We consider a link to be saturated if its ratio of saturated SFs (SFs that can transport less bits than what is queued) over allocated SFs is higher than a specified value.𝑆𝐹𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢] is a simple estimator that aims to catch the relative needs of extra SFs among e2NBs.

Finally, eq. (3) details how it is obtained for a node 𝑢. 𝑆𝐹_{𝑒𝑙𝑎𝑠𝑡𝑖𝑐}[𝑢] =

⌈ 𝑈[𝑢]

𝑁_{𝑓𝑟𝑒𝑒𝑆𝐹} ×∑

𝑢𝑈[𝑢]

⌉

(3a)

𝑈[𝑢] = ∑

𝑣∈𝑁𝑠𝑎𝑡(𝑢)

𝑇 𝐵𝑆_{𝑃 𝑅𝐵}(𝑢, 𝑣) (3b)

𝑁𝑓𝑟𝑒𝑒𝑆𝐹 = 𝐿𝑆𝐹

10×𝑆𝐹_{𝑀𝐵𝑆𝐹 𝑁} −

∑

𝑢𝑆𝐹_{𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒}[𝑢]

𝑁_𝑏/𝑁_𝑢 (3c) where 𝑈[𝑢] represents the sum of expected transport blocks per PRB among all outgoing links from node𝑢, where𝑁𝑠𝑎𝑡(𝑢) comprises the neighboring nodes of 𝑢 with saturated link along the direction from 𝑢, 𝑇 𝐵𝑆𝑃 𝑅𝐵(𝑢, 𝑣) is the expected transport block size (TBS) per allocated PRB over the link 𝑢 → 𝑣, 𝑁𝑏 is the total number of e2NB while 𝑁𝑢 is the average number of adjacent neighbors of each e2NB. Note the

Algorithm 2: Centralized Node Scheduler (centralized NS) Input : 𝑀𝐵𝑆𝐹 𝑁 is the list of MBSFN SFs during the

scheduling period𝐿𝑆𝐹.

𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡is the list of e2NBs in the network.

𝑁[𝑢]is the list of neighboring e2NBs of e2NB𝑢. 𝐿𝐼𝑁𝐾(𝑢, 𝑣)is a link between two e2NBs where e2NB𝑢acts as an eNB and e2NB𝑣as a vUE.

𝑆𝐹 𝑅𝑋[𝑢]is the list of available RX SFs at𝑢.

𝑆𝐹 𝑇 𝑋[𝑢][𝑣]is the list of available TX SFs at𝑢 for each connected e2NB𝑣.

Output:𝑆𝐹 𝑇 𝑋,𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑 foreach𝑢∈𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡do

𝑆𝐹 𝑅𝑋[𝑢] =𝑆𝐹 𝑇 𝑋[𝑢][𝑣] =𝑀𝐵𝑆𝐹 𝑁,∀𝑣. ℎ𝑎𝑣𝑒𝑂𝑛𝑒[𝑢] = 0. 𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢] =𝑆𝐹𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢].

𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢] =𝑆𝐹𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢]

𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑= 0 foreach𝑆𝐹 ∈𝑀𝐵𝑆𝐹 𝑁 do

𝑠𝑜𝑟𝑡 𝑑𝑒𝑠𝑐𝑒𝑛𝑑𝑖𝑛𝑔(𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡, 𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒)

𝑠𝑜𝑟𝑡 𝑑𝑒𝑠𝑐𝑒𝑛𝑑𝑖𝑛𝑔(𝑢∈𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡∣𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢] ==

0, 𝑒𝑙𝑎𝑠𝑡𝑖𝑐)

foreach𝑢∈𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡do

if𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢]>1or𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢]>1then if𝑆𝐹 ∈𝑆𝐹 𝑇 𝑋[𝑢][∗]then

𝑇 𝑟𝑎𝑛𝑠𝑚𝑖𝑡= 0

foreach𝑣such that𝐿𝐼𝑁𝐾(𝑢, 𝑣) existdo if𝑆𝐹 ∈𝑆𝐹 𝑅𝑋[𝑣]then

𝑇 𝑟𝑎𝑛𝑠𝑚𝑖𝑡= 1

𝑟𝑒𝑚𝑜𝑣𝑒(𝑆𝐹, 𝑆𝐹 𝑅𝑋[𝑢]) 𝑟𝑒𝑚𝑜𝑣𝑒(𝑆𝐹, 𝑆𝐹 𝑇 𝑋[𝑣][∗]) foreach𝑤∈𝑁𝑖𝑛𝑡𝑒𝑟𝑓𝑒𝑟𝑒𝑛𝑐𝑒[𝑢]do

𝑟𝑒𝑚𝑜𝑣𝑒(𝑆𝐹, 𝑆𝐹 𝑅𝑋[𝑤]) foreach𝑤∈𝑁𝑖𝑛𝑡𝑒𝑟𝑓𝑒𝑟𝑒𝑛𝑐𝑒[𝑣]do

𝑟𝑒𝑚𝑜𝑣𝑒(𝑆𝐹, 𝑆𝐹 𝑇 𝑋[𝑤][∗]) else𝑟𝑒𝑚𝑜𝑣𝑒(𝑆𝐹, 𝑆𝐹 𝑇 𝑋[𝑢][𝑣])

if𝑇 𝑟𝑎𝑛𝑠𝑚𝑖𝑡== 1then ℎ𝑎𝑣𝑒𝑂𝑛𝑒[𝑢] = 1 if𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢]>1then

𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢]− − else𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢]− −

if∀𝑢 𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒[𝑢] == 0then 𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑= 1 if∀𝑢 𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢] == 0then

foreach𝑢∈𝑒2𝑁𝐵 𝐿𝑖𝑠𝑡do 𝑒𝑙𝑎𝑠𝑡𝑖𝑐[𝑢] = 1

𝑁_{𝑓𝑟𝑒𝑒𝑆𝐹} represents the potential number of free SFs within a superframe duration (cf. eq. (1)) after allocating𝑆𝐹_{𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒}.

Algorithm 2 is used to allocate SFs between the e2NBs.

𝑆𝐹 𝑇 𝑋[𝑢][𝑣]is the list of SF on which e2NB𝑢can transmit to e2NB 𝑣 (i.e. 𝑣 will be in reception). 𝑁𝑒𝑒𝑑𝑀𝑎𝑡𝑐ℎ𝑒𝑑 indicates if the scheduler is able to give at least𝑆𝐹𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑[𝑢]

SF(s) to e2NB 𝑢 for all e2NBs or not. 𝑁𝑖𝑛𝑡𝑒𝑟𝑓𝑒𝑟𝑒𝑛𝑐𝑒(𝑢) represents the interference area around node 𝑢 and can be adjusted to contain some neighboring nodes within the radio coverage depending on the scheduling policy (e.g. conservative for a reliable transmission). In either case, the link adaption shall cope with the induced interference at the local scheduler.

(5)

Algorithm 3:Local Link scheduler (local LS) Input : 𝑢is the current e2NB identifier.

𝑆𝐹 is the current subframe identifier.

𝑈𝐸 𝐿𝑖𝑠𝑡and𝑣𝑈𝐸 𝐿𝑖𝑠𝑡are the list of UEs/vUEs with non empty queues at𝑢.

𝑄[𝑥][𝑝]is the queue size of UE/vUE𝑥for flows of priority𝑝.

𝑁𝑃 𝑅𝐵 is the number of available PRBs in𝑆𝐹. 𝑆𝐹 𝑇 𝑋[𝑢][𝑥]is the list of available Tx SFs at𝑢 for each𝑥∈𝑈𝐸 𝐿𝑖𝑠𝑡∪𝑣𝑈𝐸 𝐿𝑖𝑠𝑡.

Output:PRBs allocated for UEs and vUEs of e2NB𝑢 Result:𝑢uses𝑃 𝑅𝐵[𝑥]to transmit to each

𝑥∈𝑈𝐸 𝐿𝑖𝑠𝑡∪𝑣𝑈𝐸 𝐿𝑖𝑠𝑡 𝑠𝑜𝑟𝑡 𝑑𝑒𝑠𝑐𝑒𝑛𝑑𝑖𝑛𝑔(𝑈𝐸 𝐿𝑖𝑠𝑡, 𝑄[∗][0]) 𝑠𝑜𝑟𝑡 𝑑𝑒𝑠𝑐𝑒𝑛𝑑𝑖𝑛𝑔(𝑣𝑈𝐸 𝐿𝑖𝑠𝑡, 𝑄[∗][0]) 𝐿𝑖𝑠𝑡=𝑎𝑝𝑝𝑒𝑛𝑑(𝑣𝑈𝐸 𝐿𝑖𝑠𝑡, 𝑈𝐸 𝐿𝑖𝑠𝑡) 𝑝𝑟𝑖𝑜= 0/* 0: real-time, 1: elastic */

while𝑁𝑃 𝑅𝐵>0and𝑝𝑟𝑖𝑜 <2do 𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒 𝑝𝑟𝑖𝑜= 1

foreach𝑥∈𝐿𝑖𝑠𝑡do

if𝑆𝐹 ∈𝑆𝐹 𝑇 𝑋[𝑢][𝑥]then

if𝑁𝑃 𝑅𝐵>0∧𝑇 𝐵𝑆[𝑥]<∑_{𝑝𝑟𝑖𝑜}

𝑏=0𝑄[𝑥][𝑏]then 𝑃 𝑅𝐵[𝑥] + +

𝑁𝑃 𝑅𝐵− −

𝑇 𝐵𝑆[𝑥] =𝑇 𝐵𝑆(𝑚𝑐𝑠[𝑥], 𝑃 𝑅𝐵[𝑥]) 𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒 𝑝𝑟𝑖𝑜= 0

if𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒 𝑝𝑟𝑖𝑜== 1then 𝑝𝑟𝑖𝑜+ +

2) Local scheduling: The local scheduler is priority-based dynamic round robin algorithm on per SF basis (see Algo- rithm 3) which takes into accounts the buffer status, priorities between flows, and the channel quality indicators. It firstly allocates resources for the backhaul links (i.e. vUEs) followed by the access links (i.e. UEs) for real-time flows. The remain- ing resources will be allocated to the elastic flows until all buffers become empty or there is no PRB left.

D. Local Routing

Within the proposed algorithm, we aim to provide DL SFs at each node. It means each node can reach all its neighboring nodes using DL, and can also be reached by these neighboring nodes using UL with the associated UL SF. Hence, every two adjacent nodes can reach each other in both UL and DL directions to each other. It also means that each incoming packet can go over the DL or UL direction toward the next hop.

To this end, we use the expected waiting time of both UL and DL queues as the metric to decide whether the newly incoming packet shall go into the DL or UL queue to be delivered to the next hop.

E. Interference measurement and link adaptation

The induced interferences are multi-facet: between neighboring e2NBs happening during the backhaul SFs, between e2NBs at UE in DL, and between UEs at e2NB side in UL during the access SFs. They are caused mainly due to the neighboring e2NBs being unaware of the resource allocation policy of each other. There are two ways to deal with it. First of all, interference can be handled locally by each e2NB, simply

by using AMC to adapt the MCS of transmissions based on the reported CQIs (for DL) and its own measurement (for UL) as done by legacy systems. However, it can be improved for the backhaul link by taking into account the fact that the interference power will most probably be at the same level during the repeated SFs within the superframe as the other scheduled nodes will be the same.

Secondly, it can be handled through cooperation between nodes to achieve a better performance by dynamically enforcing the resource allocation policy based on the eNB status information. This is in contrast to (semi-)statically resource allocation policy with no or little interactions among eNBs (as in the case of X2 signaling to support inter-cell interference coordination (ICIC)). Through self-backhauling capability, eNBs can share the status information with the COE and neighboring eNBs including:

∙ the DL channel quality indicators (CQI), frequency and time resources;

∙ the UL interference status for each frequency resource (RB) experienced by the neighboring eNBs.

These above information allows to adapt the MCS, frequency resources, and the transmit power for both legacy UEs and newly-added vUEs so as to coordinate the resources and transmissions across access and backhaul links. For instance, an e2NB A is using high Tx power on frequencyresources 𝑓1 to transmit to e2NB B. With COE, e2NB C allocates a different frequency resources𝑓2to transmit to e2NB D which suffers less interference from e2NB A.

V. SIMULATIONRESULTS

This section provides the performance evaluation of the proposed coordinated scheduling and compares the results to a classical mesh network link scheduling algorithm in [11].

A. Simulation environment

A complete LTE simulator is developed in MATLAB allowing to create a 2D-map of an arbitrary network of e2NBs with their associated UEs and to generate arbitrary flows between every type of nodes (e.g. U2U, e2U, e2e traffic). We assume that it takes 5ms for a eNB to process an incoming RF packet before pushing it to the output DL or UL queue to be transported toward the next hop as introduced beforehand.

1) Channel models: Between e2NBs, a freespace path loss model of coefficient 2.1 is applied with Claussen shadow fading and EPA multi-path channel. Between e2NBs and UEs, a rural (TR 36.942) path loss model is used with Claussen shadow fading and EPA multi-path channel. No further assumption is made on the interference coordination and mitigation for UEs. Moreover, e2NBs and UEs are assumed to be fixed and equipped with an omni directional antenna.

All interferences caused by concurrent transmission are taken into account at each receiver.

2) Topologies: We define two different topologies as shown in Fig. 3(a) and 3(b). In both topologies, each BS has 10 attached UEs and is connected to the closest adjacent e2NBs through the co-located vUEs.

(6)

(a) Line Topology with 5 e2NBs

(a) Hexagonal Topology with 7 e2NBs

Fig. 3: Considered network topologies

3) Traffic patterns: Two traffic patterns are considered.

First of all, all UEs are randomly paired with each other and establish a bi-directional real-time VoIP call with a packet size of 20 bytes (40 bytes on PHY using RoHC) with an arrival rate of 20ms. Secondly, two fixed e2e elastic traffic flows are added. Each elastic traffic aims to maximize its throughput and behaves as a connection-oriented acknowledgment service in that the new packet will be only generated if the maximum number of non-acknowledged packets is not reached. This type of flow represents the inter-site (i.e. BSs) data transfer that often happens in military and public safety application scenarios and is not directly generated by UEs.

4) Other settings and parameters: HARQ mechanism is not implemented in the simulator. UEs are only scheduled in UL/DL transmission in non MBSFN SFs. Expired VoIP packets (one-way-delay larger than 200ms) are discarded from the queue to improve the transportation efficiency. A link is considered to be saturated if the utilization ratio is larger than 80%. Each e2NB is configured with 10MHz channel bandwidth in band 4 (2.1GHz). For the BSs, the minimum transmit power is set to 20dBm and the maximum to 40dBm for DL UE transmissions and 46dBm for inter-e2NB transmissions.

For UEs, maximum UL transmit power is set to 23dBm.

5) Li and Ephremides algorithm: In [11], Li and Ephremides propose a joint scheduling, power control and routing algorithm for TDMA wireless mesh networks. Their approach relies on a centralized scheduler that is executed at each time slot and decides which links should be activated as well as the transmission power. Their algorithm firstly sorts all possible links according to the link metric, which is based on the queue size of itself and all other links blocked by this link to prioritize the large queues while handling the congestion.

For instance, a link 𝑖is considered to be blocked by a link𝑗 if its source or destination is identical to either the source or destination of link𝑗. Then, an iterative algorithm can find the set of links with the highest priority that can be concurrently activated while respecting their power limitation and SINR requirements given all induced interference. In addition, their method periodically updates the routing paths based on the link quality and the relative queue size of each link to allow

packets being redirected towards less congested links when queue buffer occupancy increases.

One of the main drawbacks of the Li and Ephremides approach is that the link scheduling is done in per-slot basis and thus it requires a perfect knowledge of queue sizes of all nodes and channel behavior between all nodes. Moreover, each activated node is used for a single point-to-point link which is inefficient in LTE system as the multi-point to point transmissions are allowed in both DL and UL directions within the same slot.

6) Improvement on Li and Ephremides algorithm: The following improvement is implemented to be applied by the Li and Ephremides algorithm in LTE for a fair comparison.

We update the link metric and take into account both DL and UL queues of each link. Moreover, inter-node link scheduling is done for both DL and UL directions. The iterative power allocation algorithm is modified to incorporate the adaptive modulation and coding of different SINR regimes.

B. Algorithms considered

In following, we consider three mixed combinations of global and local approaches. In terms of the global approach, we consider three algorithms to be compared: (1) the Li and Ephremides algorithm with the aforementioned modifications (denoted as Li algorithm), (2) Algorithm 2 that uses the same and fixed number of MBSFN SFs between nodes for both real-time (𝑆𝐹𝑟𝑒𝑎𝑙𝑡𝑖𝑚𝑒) and elastic (𝑆𝐹𝑒𝑙𝑎𝑠𝑡𝑖𝑐) traffic among all e2NBs (denoted as the basic algorithm), and (3) the full proposed Algorithm 2 as described in section IV. Moreover, we consider two different variants on the local scheduler, (1) uses Algorithm 3 for UE scheduling but allocates all RBs of inter-e2NB links to a single vUE at one time while (2) and (3) use Algorithm 3 to allocate RBs in a fine-grained manner. In all cases, the incoming packets are classified into either real-time VoIP traffic or elastic traffic and are sorted and prioritized at the local vUE queues based on two manners:

(a) the number of hops to reach the destination, labeled as Hops in following, or (b) the shortest deadline, labeled as AirTime. These two different metrics will affect the number of dropped VoIP packets, and consequently the throughput of elastic traffic.

C. Results

Firstly, we consider two scenarios: (a) only VoIP traffic, and (b) both VoIP and elastic traffics. Then, we show the results in two forms: (i) the CDF plot in the percentage of VoIP packets that are within the pre-defined delay requirement (150ms) and (ii) throughput of elastic traffic flows.

1) Line topology: In Fig. 4.(a), it can be seen that with only VoIP traffic, all three approaches satisfy the required 95% of packet with less than 150ms end-to-end latency among all VoIP flows. This is because there is sufficient capacity to transport all flows even without using the multi-point to point characteristic of LTE, particularly because the topology allows for a maximum of two point-to-point communications (two adjacent neighbors) from a given node at a time slot.

(7)

(a) CDF of VoIP flows. (50 VoIP flows)

(b) CDF of VoIP flows. (50 VoIP and 2 e2e elastic flows)

(c) Cumulated data rate of elastic traffic. (50 VoIP and 2 e2e elastic flows)

Fig. 4: Line Topology

(a) CDF of VoIP flows. (70 VoIP flows)

(b) CDF of VoIP flows. (70 VoIP and 2 e2e elastic flows)

(c) Cumulated data rate of elastic traffic. (70 VoIP and 2 e2e elastic flows)

Fig. 5: Hexagonal topology

Then, we add two extra e2e elastics flows (5 → 4 and 2 →4). In Fig. 4.(b), we can see that Li approach performs poorly regarding latency requirements as only 60% of flows meet the requirement when using AirTimeas a sorting metric and only 40% of flows meet the requirement when using Hops as a sorting metric. On the other hand, bothBasicand Proposed approaches allows all flows to meet their latency requirement. The Basic approach is comparable with the Proposed approach and the AirTime sorting metric is better than the Hopsapproach. From Fig. 4.(c), it can be observed that the Basic approach has the lowest cumulated throughput as it fixed global allocation which does not react to the needs of elastic flows, while the Proposed approach on average is comparable with the Li approach.

2) Hexagonal topology: For the hexagonal topology, we observe in Fig. 5.(a) that around 5% of VoIP flows cannot meet the latency requirement when using Li approach while all flows underBasicandProposedapproaches can fulfill the requirement. Then, two extra e2e elastics flows (5 →4 and 2→4) are added. In the Fig. 5.(b), bothBasicandProposed approaches perform perfectly regarding latency requirements but Li approach fall short with almost 90% of the VoIP flows that do not meet the requirements. From Fig. 5.(c), Li approach performs the worst while theProposedapproach achieves 35%

throughput improvement over theBasicapproach. The under- performance of Li approach is due to its lack of multi-point to point capability that is highly beneficial in networks with multiple adjacent nodes.

To sum up, theProposedapproach achieves the best latency- throughput trade-off for real-time and elastic traffics. Such merit is mainly due to its link adaptation capability and the use of multi-point to point link scheduling.

VI. CONCLUSION

In this article, we present a method to efficiently and practically realize an in-band LTE wireless mesh networks that leverages the LTE relay interface (Un) and takes into account QoS requirements of real-time flows. We then compare such approach to two other ones in different scenarios. The findings are dual: first of all, we show that the in-band meshing of LTE BSs is possible without requiring intense modifications on current standards while keeping legacy UE support. Secondly, we show that our approach is able to guarantee QoS requirements of real-time flows while improving the network throughput for elastic flows. In future, we plan to detailedly introduce the architecture to realize the autonomy of network based on the preliminary study in [12].

REFERENCES

[1] R. Favraud et al., “Toward moving public safety networks,” IEEE Communications Magazine, vol. 54, no. 3, pp. 14–20, March 2016.

[2] R. Favraud and N. Nikaein, “Wireless mesh backhauling for lte/lte-a networks,” inIEEE MILCOM 2015, Oct 2015, pp. 695–700.

[3] ——, “Analysis of LTE Relay Interface for Self-Backhauling in LTE Mesh Networks,” 2017.

[4] Y. Yuan, LTE-Advanced Relay Technology and Standardization.

Springer-Verlag Berlin Heidelberg, 2013.

[5] A. S. Mohamedet al., “Self-organized dynamic ffr resource allocation scheme for lte-advanced relay based networks,”Wireless Personal Com- munications, vol. 91, no. 2, pp. 933–955, 2016.

[6] J. Han and H. Wang, “Uplink performance evaluation of wireless self- backhauling relay in lte-advanced,” inWiCOM 2010, Sept 2010.

[7] P. H. Pathak and R. Dutta, “A survey of network design problems and joint design approaches in wireless mesh networks,”IEEE Communica- tions Surveys Tutorials, vol. 13, no. 3, pp. 396–428, Third 2011.

[8] A. F. Ribadeneira, “An analysis of the MOS under conditions of delay, jitter and packet loss and an analysis of the impact of introducing piggybacking and reed solomon FEC for VoIP,” Master’s thesis, Georgia State University, 2007.

[9] X. Foukaset al., “Flexran: A flexible and programmable platform for software-defined radio access networks,” inACM CoNEXT 2016, 2016.

[10] I. F. Akyildizet al., “Wireless mesh networks: A survey,”Comput. Netw.

ISDN Syst., vol. 47, no. 4, pp. 445–487, Mar. 2005.

[11] Y. Li and A. Ephremides, “A joint scheduling, power control, and routing algorithm for ad hoc wireless networks,”Ad Hoc Networks, vol. 5, no. 7, pp. 959 – 973, 2007.

[12] R. Favraudet al., “Self-backhauled autonomous LTE mesh networks,”

inIEEE WIMOB, 2017.